fix: check adapter

Signed-off-by: Meow <ongjackm@gmail.com>

xlm_padding.py

@@ -98,7 +98,7 @@ class IndexFirstAxisResidual(torch.autograd.Function):
 index_first_axis_residual = IndexFirstAxisResidual.apply
 
 
-def unpad_input(hidden_states, attention_mask, adapter_mask):
+def unpad_input(hidden_states, attention_mask, adapter_mask=None):
     """
     Arguments:
         hidden_states: (batch, seqlen, ...)
@@ -114,11 +114,13 @@ def unpad_input(hidden_states, attention_mask, adapter_mask):
     max_seqlen_in_batch = seqlens_in_batch.max().item()
     cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0))
 
-    cu_adapter_mask = torch.empty(cu_seqlens[-1], dtype=torch.int32)
-    for i in range(len(adapter_mask)):
-        start_idx = cu_seqlens[i]
-        end_idx = cu_seqlens[i + 1]
-        cu_adapter_mask[start_idx:end_idx] = adapter_mask[i]
+    cu_adapter_mask = None
+    if adapter_mask:
+        cu_adapter_mask = torch.empty(cu_seqlens[-1], dtype=torch.int32)
+        for i in range(len(adapter_mask)):
+            start_idx = cu_seqlens[i]
+            end_idx = cu_seqlens[i + 1]
+            cu_adapter_mask[start_idx:end_idx] = adapter_mask[i]
 
     # TD [2022-03-04] We don't want to index with a bool mask, because Pytorch will expand the
     # bool mask, then call nonzero to get the indices, then index with those. The indices is @dim